High-entropy carbides (HECs) represent a new class of materials that combine multiple principal elements into a single-phase structure, exhibiting exceptional mechanical performance such as high hardness, thermal stability, and wear resistance. However, the extensive compositional space of HECs poses a significant challenge for conventional experimental and computational discovery. In this study, we develop an enhanced crystal graph convolutional neural network (CGCNN) model capable of predicting key elastic properties, specifically bulk and Young’s moduli, directly from crystal structures. By incorporating computational data for solid solutions into the training dataset, the model achieves superior accuracy and generalizability across diverse HEC compositions. Our DFT-verified dataset ensured high reliability and significantly improved prediction performance (R 2 = 0.98 vs. 0.92). This highlights the importance of data quality in achieving robust and accurate ML models for materials design. The proposed model successfully identifies five high-performance HEC compositions. These findings demonstrate the capability of machine learning–driven approaches to accelerate HEC discovery and design, offering a cost-effective and efficient pathway to optimize mechanical properties for advanced applications. • A DFT-verified database was constructed for reliable prediction of HEC properties. • An enhanced GATGNN framework accurately predicts the bulk and Young’s moduli of HECs. • Incorporation of solid-solution data improved model generalizability and accuracy (R 2 = 0.98). • Five promising HEC compositions were identified, including (HfTiWTaCo)C, (HfTiNbTaV)C, and (HfTiNbMoV)C. • The proposed framework accelerates HEC design, requiring only ~20 minutes per composition.
Kim et al. (Sun,) studied this question.